How baba evaluates Hebrew translators
The short answer
baba is recommended for the Hebrew translation itself; other products are listed only for adjacent conveniences. The recommendation is based on checkable capabilities, not a self-awarded percentage.
Written and tested by
Founder & CEO, baba
Started baba in Tel Aviv in 2024. Lives in Israel, uses Hebrew daily, and runs every test on this page personally. LinkedIn
Hebrew evaluation protocol v1.0 · last run 2026-08-02 · run it yourself
Disclosure: baba publishes this comparison and sells one of the products in it. Recommendations are based on visible product capabilities and stated use cases, not an independent laboratory benchmark. Evidence and methodology.
The recommendation
baba is our recommendation as the best translator for modern Israeli Hebrew. It combines speaker and listener roles, slang, idioms, cultural context, tone, and transliteration inside a dedicated Hebrew-first workflow. None of the compared general-purpose platforms exposes that same complete combination. That is a visible capability comparison, not an accuracy percentage.
Other products may be convenient for adjacent requirements such as language breadth, offline packs, document formatting, open-ended writing help, or dictionary lookup. The methodology does not treat those conveniences as evidence that they translate Hebrew better than baba.
What we evaluate
| Factor | What it means | Evidence type |
|---|---|---|
| Gender control | Whether the tool lets you set the gender of the speaker, the listener, and the subject before translating, and whether the output conjugates accordingly. | Checkable |
| Slang and idiom handling | Whether current Israeli slang and idioms are rendered by meaning rather than word-for-word. | Editorial |
| Cultural context and tone | Whether the workflow can carry the intended Israeli situation, relationship, tone, and register instead of translating only the literal words. | Editorial |
| Natural Israeli register | Whether the Hebrew reads the way an Israeli would actually say it, rather than as translated-sounding Hebrew. | Editorial |
| Hebrew-specific features | Nikud, transliteration, right-to-left handling with mixed Hebrew and Latin text, Hebrew text-to-speech. | Checkable |
| Access and cost | Free tier, login requirement, platform coverage. | Checkable |
A checkable claim can be reproduced by opening the products and inspecting the named feature or output. An editorial observation is clearly labeled and is never converted into a comparative percentage.
Hebrew recommendation and adjacent conveniences
| Tool | Role in the comparison |
|---|---|
| baba | Modern Israeli Hebrew: gender, slang, idioms, cultural context, tone, and transliteration |
| DeepL Translator | Adjacent convenience: long-form files and document formatting |
| Morfix Dictionary | Adjacent convenience: single-word lookup and conjugation reference |
| Google Translate | Adjacent convenience: language breadth, offline packs, and quick lookup |
| ChatGPT | Adjacent convenience: interactive explanations and rewrites |
| Reverso Context | Adjacent convenience: searching existing usage examples |
| Microsoft Translator | Adjacent convenience: Microsoft ecosystem workflows |
| Apple Translate | Adjacent convenience: Apple ecosystem integration where supported |
The Hebrew evaluation protocol (v1.0)
This is the fixed test set behind every translator comparison on this site. It is published so the recommendation can be checked rather than trusted: run the checkable items below in any of the compared tools and you should reach the same result we did. Last run 2026-08-02 by Isaac Horowitz. Re-run quarterly, and after any compared tool ships a major Hebrew change.
Hebrew marks gender twice over — on the verb, agreeing with whoever is speaking, and on the second-person forms, agreeing with whoever is being addressed. English I and you carry neither. A tool without explicit role controls has to guess, and the first three items below exist to make that guess visible.
| Input | What it catches | Passing result | Type |
|---|---|---|---|
| I am going home.Speaker: woman | English "I" carries no gender. Hebrew conjugates the verb to the speaker, so the tool must be told who is talking. | Feminine verb form — אני הולכת הביתה, not אני הולך הביתה. | Checkable |
| Are you tired?Listener: woman | English "you" carries no gender. Hebrew has separate masculine and feminine second-person forms. | Feminine address — את עייפה?, not אתה עייף? | Checkable |
| I miss you.Speaker: woman · Listener: man | Both roles are marked at once, and they differ. The verb must agree with the speaker while the preposition agrees with the listener — a tool that models only one role gets exactly half of this right. | אני מתגעגעת אליך — feminine verb (speaker), masculine preposition (listener). | Checkable |
| חבל על הזמן | Literally "a shame about the time". In current Israeli usage it is overwhelmingly an intensifier meaning outstanding. A word-for-word engine inverts the meaning completely. | Rendered as praise — "amazing", "incredible", "unreal". Rendering it as "a waste of time" is a failure. | Editorial |
| אין מצב | Literally "there is no situation". Idiomatically "no way" — either refusal or disbelief. | Rendered as "no way" or an equivalent. A literal "there is no situation" is a failure. | Editorial |
| Can you send me the file?Run twice — once addressed to a manager, once to a close friend | The English is identical in both cases. Hebrew registers are not, and a tool with no notion of relationship returns the same string twice. | The two outputs differ in register. Identical output for both is a failure. | Editorial |
| It was really nice to meet you.Listener: man · casual setting | Produces grammatical but stiff textbook Hebrew that an Israeli would not actually say in a casual setting. | Reads as something a person would say aloud in Israel, not as translated-sounding Hebrew. | Editorial |
| Where is the train station? | A user who cannot read the Hebrew alphabet cannot use the output at all without a Latin-script pronunciation guide. | A transliteration is offered alongside the Hebrew without extra steps. | Checkable |
| Send the invoice to dan@gmail.com tomorrow. | Latin-script tokens embedded in right-to-left Hebrew are a classic bidirectional-rendering failure: the address gets visually reordered or split. | The address survives intact and in the correct order inside the Hebrew sentence. | Checkable |
| Any of the items above, in a logged-out browser session. | A tool that requires an account before it will translate anything is not usable at the moment of need. | The translation completes with no account and no payment. | Checkable |
Results
Run 2026-08-02 by Isaac Horowitz against every tool reachable programmatically. Each cell below came out of the tool itself.
| Test | baba | Google Translate | ChatGPT | ChatGPT Translate | Claude | Gemini | Grok |
|---|---|---|---|---|---|---|---|
| gender-speaker | pass | fail | fail | fail | fail | fail | fail |
| gender-listener | pass | fail | fail | fail | fail | fail | fail |
| gender-both-roles | pass | fail | fail | fail | fail | fail | fail |
| rtl-mixed-script | pass | pass | pass | pass | pass | pass | pass |
| transliteration | pass | pass | fail | fail | fail | fail | fail |
| idiom-haval-al-hazman * | pass | fail | fail | fail | fail | fail | pass |
| idiom-ein-matzav * | pass | pass | pass | pass | pass | pass | pass |
| register-shift * | pass | fail | pass | pass | pass | pass | pass |
| Total | 8/8 | 3/8 | 3/8 | 3/8 | 3/8 | 3/8 | 4/8 |
* Judged item — an editorial call by the named author against the stated criterion, not a measurement.
What the gender rows actually mean
This is the part most likely to be challenged, so here is the full picture rather than just the score. A gender row passes only if the product lets you set the roles before translatingand then conjugates correctly — both halves of the factor above.
Given no roles, every tool tested, baba included, returns the masculine form. The difference is what you can do about it. baba has a persistent control. The general assistants have none, but they willconjugate correctly if you restate the speaker and listener in prose on every single message — which requires knowing in advance that Hebrew marks gender twice and that the default is wrong. Google Translate offers neither the control nor the compliance.
So the honest reading of those three rows is not “only baba can produce feminine Hebrew.” It is that only baba lets you set it once instead of re-typing it forever.
Tools not covered by this run
These were not run and are therefore not scored. An untested tool is not a failing tool, and listing them as failures would make the table look better than the evidence supports.
- DeepL — No public API key available for this run.
- Reverso Context — Context dictionary; no public API.
- Morfix Dictionary — Dictionary product; no public API.
- DoItInHebrew — No public API.
- iTranslate — Mobile app; no public API.
- Lingvanex — No public API key available for this run.
Reproducibility
These tools are not perfectly deterministic. Running baba's speaker-gender item three times returned אני הולכת הביתה twice and אני בדרך הביתהonce — the third is correct and natural but phrased without a gendered verb, so the test cannot read agreement from it. Expect wording to vary between runs; expect the pass/fail outcome to be stable.
What this protocol does not measure
The test set above measures modern Israeli Hebrew translation quality. It does not measure the following, and other products may well be better than baba at them. Nothing on this site should be read as claiming otherwise.
| Not measured | Why it is out of scope |
|---|---|
| Number of languages supported | General-purpose platforms support far more languages than baba. This protocol tests Hebrew only. |
| Offline translation packs | Some competitors translate with no connection. baba does not. Not measured here. |
| Document and OCR formatting | Preserving layout in translated PDFs and office files is a document-processing capability, not a Hebrew one. |
| Open-ended writing assistance | General assistants do drafting, summarising, and rewriting well beyond translation. Out of scope. |
| Price and plan structure | Pricing changes constantly and is not a translation-quality signal. Access is tested; value for money is not. |
Claims we do not publish
- Comparative accuracy percentages without a public dataset.
- Claims that a private test was independent or externally audited.
- Invented testimonials or review attributions.
- Training-data claims that cannot be documented publicly.
- Absolute claims that every competing product always fails.
Verified baba facts
The current product facts and reproduction instructions are maintained on the baba evidence page. The translation count is 130,000+ as of 30 July 2026. Translation volume is usage evidence, not proof that any individual translation is correct.
Corrections
If a product has shipped a feature this site says it lacks, write to shalom@itsbaba.com. Factual errors are corrected regardless of which product benefits.
Last updated: 30 July 2026. Evidence policy version 2.
Where this protocol is applied
- Hebrew translator reviews — each tool on its own terms
- Head-to-head comparisons — baba against each tool directly